OpenAI spending hits $34 billion in 2025: Ed Zitron

William Foxley

Tue, June 16, 2026 astatine 7:29 AM CDT 2 min read

Audited fiscal documents reviewed by Ed Zitron's Where's Your Ed At and independently shown to the Financial Times amusement OpenAI is burning billions of dollars annually.

Per Zitron, OpenAI posted $13.07 cardinal successful 2025 gross against $34 cardinal successful full costs and expenses, producing a nett nonaccomplishment attributable to the institution of $38.53 billion. That nonaccomplishment is astir 7.5 times the $5.09 cardinal it mislaid successful 2024 connected $3.7 cardinal successful revenue.

"The fiscal information of OpenAI is profoundly concerning. $38.53 cardinal successful losses are astronomical, and acold higher than astir believed it would be," Zitron said successful the blog post. "Losses besides look to beryllium mounting year-over-year astatine a melodramatic rate, and I'm not definite however this institution finds a mode toward immoderate benignant of sustainability oregon profitability."

OpenAI declined to remark to the FT.

Where the wealth goes

The documents interruption costs into categories that uncover however capital-intensive frontier inference has become. Cost of revenue, the enactment that captures serving exemplary outputs to users, roseate from $2.65 cardinal successful 2024 to $7.5 cardinal successful 2025. Research and improvement jumped from $7.81 cardinal to $19.18 billion. Sales and selling went from $1.11 cardinal to $5.73 billion.

OpenAI spent $5.02 cardinal connected inference with Microsoft Azure successful the archetypal fractional of 2025 alone. For the afloat span from calendar twelvemonth 2024 done Q3 2025, inference walk connected Azure totaled $12.43 billion, per the documents. That fig covers lone inference, not training, which means the afloat compute measure is larger still.

OpenAI's apical enactment is increasing fast, but its outgo basal is increasing faster due to the fact that each new, much susceptible exemplary demands much compute to bid and much compute to serve. Scaling the merchandise means scaling the infrastructure underneath it, and that infrastructure runs connected GPUs, power, and data-center capableness that bash not get cheaper arsenic request rises.